P_Ctx P = Zi;
ThreadLocal P_ThreadLocalCtx P_tl = Zi;

Readonly P_Ent P_NilEnt = {
  .xf = CompXformIdentity,
  .look = { 0, -1 },
};

Readonly P_Frame P_NilFrame = {
  .first_ent = &P_NilEnt,
  .last_ent = &P_NilEnt,
};

////////////////////////////////////////////////////////////
//~ Bootstrap

void P_Bootstrap(void)
{
  // Initialize shared state
  for (u64 i = 0; i < countof(P.sim_input_states); ++i)
  {
    P_InputState *input = &P.sim_input_states[i];
    input->arena = AcquireArena(Gibi(64));
  }
  for (u64 i = 0; i < countof(P.sim_output_states); ++i)
  {
    P_OutputState *output = &P.sim_output_states[i];
    output->arena = AcquireArena(Gibi(64));
  }
}

////////////////////////////////////////////////////////////
//~ Nil helpers

b32 P_IsKeyNil(P_Key key)
{
  return key.v == 0;
}

b32 P_IsEntNil(P_Ent *ent)
{
  return ent == 0 || ent == &P_NilEnt;
}

b32 P_IsFrameNil(P_Frame *frame)
{
  return frame == 0 || frame == &P_NilFrame;
}

////////////////////////////////////////////////////////////
//~ Key helpers

b32 P_MatchKey(P_Key a, P_Key b)
{
  return a.v == b.v;
}

P_Key P_RandKey(void)
{
  // TODO: Don't use true randomness for entity keys. It's overkill & non-deterministic.
  P_Key result = Zi;
  TrueRand(StringFromStruct(&result));
  return result;
}

////////////////////////////////////////////////////////////
//~ Tile helpers

String P_NameFromTileKind(P_TileKind kind)
{
  // Tile names array
#define X(name, ...) [P_TileKind_##name] = CompLit(#name),
  PERSIST Readonly String tile_names[] = {
    P_TilesXMacro(X)
  };
#undef X

  String result = Lit("Unknown");
  if (kind >= 0 && kind < countof(tile_names))
  {
    result = tile_names[kind];
  }
  return result;
}


////////////////////////////////////////////////////////////
//~ Shape helpers

P_Shape P_ShapeFromDescEx(P_ShapeDesc desc)
{
  desc.count = MaxI32(desc.count, 1);
  P_Shape result = Zi;
  {
    result.points_count = desc.count;
    CopyStructs(result.points, desc.points, result.points_count);
    Vec2 accum = Zi;
    for (i32 p_idx = 0; p_idx < result.points_count; ++p_idx)
    {
      accum = AddVec2(accum, result.points[p_idx]);
    }
    result.centroid = DivVec2(accum, result.points_count);
    result.center_of_mass = result.centroid;
    result.radius = desc.radius;
    result.mass = desc.mass;
  }
  return result;
}

P_Shape P_MulXformShape(Xform xf, P_Shape shape)
{
  P_Shape result = shape;
  for (i32 i = 0; i < shape.points_count; ++i)
  {
    result.points[i] = MulXformV2(xf, shape.points[i]);
  }
  Vec2 scale = ScaleFromXform(xf);
  result.radius *= MaxF32(scale.x, scale.y);
  result.centroid = MulXformV2(xf, shape.centroid);
  result.center_of_mass = MulXformV2(xf, shape.center_of_mass);
  return result;
}

Rng2 P_BoundingBoxFromShape(P_Shape shape)
{
  Vec2 left = P_SupportPointFromShape(shape, VEC2(-1, 0)).p;
  Vec2 top = P_SupportPointFromShape(shape, VEC2(0, -1)).p;
  Vec2 right = P_SupportPointFromShape(shape, VEC2(1, 0)).p;
  Vec2 bottom = P_SupportPointFromShape(shape, VEC2(0, 1)).p;

  Rng2 result = Zi;
  result.p0 = VEC2(left.x, top.y);
  result.p1 = VEC2(right.x, bottom.y);
  return result;
}

P_Shape P_LocalShapeFromEnt(P_Ent *ent)
{
  P_Shape result = Zi;

  // TODO: This is a temporary hack. We should eventually switch to using a prefab lookup table.
  if (ent->is_player)
  {
    result = P_ShapeFromDesc(
      .mass = 10,
      .count = 1,
      .radius = 0.3,
    );

    // f32 player_width = 0.6;
    // f32 player_height = 0.3;
    // result = P_ShapeFromDesc(
    //   .mass = 10,
    //   .count = 2,
    //   .points = { VEC2(-player_width / 2 + (player_height / 2), 0), VEC2(player_width / 2 - (player_height / 2), 0) },
    //   .radius = player_height / 2,
    // );

    // Rng2 test_rect = Zi;
    // test_rect.p0 = VEC2(-1, -1);
    // test_rect.p1 = VEC2(1, 1);
    // result = P_ShapeFromDesc(
    //   // .radius = 0.5,
    //   .radius = 0,
    //   .count = 4,
    //   .points[0] = VEC2(test_rect.p0.x, test_rect.p0.y),
    //   .points[1] = VEC2(test_rect.p1.x, test_rect.p0.y),
    //   .points[2] = VEC2(test_rect.p1.x, test_rect.p1.y),
    //   .points[3] = VEC2(test_rect.p0.x, test_rect.p1.y),
    // );
  }

  return result;
}

P_Shape P_WorldShapeFromEnt(P_Ent *ent)
{
  P_Shape local = P_LocalShapeFromEnt(ent);
  P_Shape world = P_MulXformShape(ent->xf, local);
  return world;
}

////////////////////////////////////////////////////////////
//~ Collision

// NOTE: Everything here is pretty much copied directly from the old prototype.
// The techniques are slow and do more than what we need. For example we should
// probably just switch from GJK to SAT for shape collision testing.

P_SupportPoint P_SupportPointFromShapeEx(P_Shape shape, Vec2 dir, i32 ignore_idx)
{
  P_SupportPoint result = Zi;
  Vec2 dir_norm = NormVec2(dir);
  f32 max_dot = -Inf;
  if (shape.points_count == 1)
  {
    // Don't ignore for single-point colliders
    ignore_idx = -1;
  }
  for (i32 point_idx = 0; point_idx < shape.points_count; ++point_idx)
  {
    if (point_idx != ignore_idx)
    {
      Vec2 p = shape.points[point_idx];
      f32 dot = DotVec2(p, dir_norm);
      if (dot > max_dot)
      {
        max_dot = dot;
        result.p = p;
        result.id = point_idx;
      }
    }
  }
  result.p = AddVec2(result.p, MulVec2(dir_norm, shape.radius));
  return result;
}

P_SupportPoint P_SupportPointFromShape(P_Shape shape, Vec2 dir)
{
  return P_SupportPointFromShapeEx(shape, dir, -1);
}

P_MenkowskiPoint P_MenkowskiPointFromShapes(P_Shape shape0, P_Shape shape1, Vec2 dir)
{
  P_MenkowskiPoint result = Zi;
  result.s0 = P_SupportPointFromShape(shape0, dir);
  result.s1 = P_SupportPointFromShape(shape1, NegVec2(dir));
  result.p = SubVec2(result.s0.p, result.s1.p);
  return result;
}

P_ClippedLine P_ClipLineToLine(Vec2 a0, Vec2 b0, Vec2 a1, Vec2 b1, Vec2 normal)
{
  Vec2 vab0 = SubVec2(b0, a0);
  Vec2 vab1 = SubVec2(b1, a1);
  Vec2 va0a1 = SubVec2(a1, a0);
  Vec2 vb0b1 = SubVec2(b1, b0);
  f32 vab0_w = WedgeVec2(vab0, normal);
  f32 vab1_w = WedgeVec2(vab1, normal);
  f32 va0a1_w = WedgeVec2(va0a1, normal);
  f32 vb0b1_w = WedgeVec2(vb0b1, normal);

  // FIXME: Handle 0 denominator
  f32 a0t;
  f32 b0t;
  {
    f32 w = 1 / vab0_w;
    a0t = ClampF32(va0a1_w * w, 0, 1);
    b0t = ClampF32(vb0b1_w * -w, 0, 1);
  }
  f32 a1t;
  f32 b1t;
  {
    f32 w = 1 / vab1_w;
    a1t = ClampF32(-va0a1_w * w, 0, 1);
    b1t = ClampF32(-vb0b1_w * -w, 0, 1);
  }

  P_ClippedLine result = Zi;
  result.a0_clipped = AddVec2(a0, MulVec2(vab0, a0t));
  result.a1_clipped = AddVec2(a1, MulVec2(vab1, a1t));
  result.b0_clipped = AddVec2(b0, MulVec2(vab0, -b0t));
  result.b1_clipped = AddVec2(b1, MulVec2(vab1, -b1t));
  return result;
}

Vec2 P_ClipPointToLine(Vec2 a, Vec2 b, Vec2 p, Vec2 normal)
{
  Vec2 vab = SubVec2(b, a);
  Vec2 vap = SubVec2(p, a);

  f32 vab_w = WedgeVec2(vab, normal);
  f32 vap_w = WedgeVec2(vap, normal);

  f32 w = 1 / vab_w;
  f32 t = ClampF32(vap_w * w, 0, 1);

  Vec2 result = AddVec2(a, MulVec2(vab, t));
  return result;
}

P_CollisionResult P_CollisionResultFromShapes(P_Shape shape0, P_Shape shape1)
{
  P_CollisionResult result = Zi;
  TempArena scratch = BeginScratchNoConflict();

  f32 tolerance = 0.00005f;  // How close can non-overlapping shapes be before collision is considered
  f32 min_unique_pt_dist_sq = (0.001f * 0.001f);  // NOTE: Should always be less than tolerance, since colliding = 1 if origin is within this distance.
  u32 max_iterations = 64;  // To prevent extremely large prototypes when origin is in exact center of rounded feature

  //////////////////////////////
  //- GJK

  P_MenkowskiSimplex simplex = Zi;
  Vec2 non_overlapping_dir = Zi;
  b32 is_overlapping = 0;
  {
    P_MenkowskiPoint m = Zi;

    // First point is support point in shape's general directions to eachother
    Vec2 dir = SubVec2(shape1.centroid, shape0.centroid);
    if (IsVec2Zero(dir)) dir = VEC2(1, 0);
    simplex.a = P_MenkowskiPointFromShapes(shape0, shape1, dir);
    simplex.count = 1;

    Vec2 removed_a = Zi;
    Vec2 removed_b = Zi;
    u32 num_removed = 0;
    for (;;)
    {
      //////////////////////////////
      //- Find initial points in simplex

      if (simplex.count == 1)
      {
        // Second point is support point towards origin
        dir = NegVec2(simplex.a.p);

        m = P_MenkowskiPointFromShapes(shape0, shape1, dir);
        // Check that new point is far enough away from existing point
        if (Vec2LenSq(SubVec2(m.p, simplex.a.p)) < min_unique_pt_dist_sq)
        {
          is_overlapping = 0;
          break;
        }
        simplex.b = simplex.a;
        simplex.a = m;
        simplex.count = 2;

        // Third point is support point in direction of line normal towards origin
        dir = PerpVec2TowardsDir(SubVec2(simplex.b.p, simplex.a.p), NegVec2(simplex.a.p));
      }

      //////////////////////////////
      //- Find third point in simplex

      {
        m = P_MenkowskiPointFromShapes(shape0, shape1, dir);
        // Check that new point is far enough away from existing points
        if (
          Vec2LenSq(SubVec2(m.p, simplex.a.p)) < min_unique_pt_dist_sq ||
          Vec2LenSq(SubVec2(m.p, simplex.b.p)) < min_unique_pt_dist_sq || (
            (num_removed >= 1) && (
              (Vec2LenSq(SubVec2(m.p, removed_a)) < min_unique_pt_dist_sq) ||
              (num_removed >= 2 && Vec2LenSq(SubVec2(m.p, removed_b)) < min_unique_pt_dist_sq)
            )
          ) ||
          AbsF32(WedgeVec2(SubVec2(simplex.b.p, simplex.a.p), SubVec2(m.p, simplex.a.p))) < min_unique_pt_dist_sq
        )
        {
          is_overlapping = 0;
          break;
        }
        simplex.c = simplex.b;
        simplex.b = simplex.a;
        simplex.a = m;
        simplex.count = 3;

        if (
          (AbsF32(WedgeVec2(SubVec2(simplex.b.p, simplex.a.p), NegVec2(simplex.a.p))) <= min_unique_pt_dist_sq) ||
          (AbsF32(WedgeVec2(SubVec2(simplex.c.p, simplex.b.p), NegVec2(simplex.b.p))) <= min_unique_pt_dist_sq) ||
          (AbsF32(WedgeVec2(SubVec2(simplex.c.p, simplex.a.p), NegVec2(simplex.a.p))) <= min_unique_pt_dist_sq)
        )
        {
          // Simplex lies on origin
          is_overlapping = 1;
          break;
        }
      }

      //////////////////////////////
      //- Determine origin region

      Vec2 vab = SubVec2(simplex.b.p, simplex.a.p);
      Vec2 vac = SubVec2(simplex.c.p, simplex.a.p);
      Vec2 vbc = SubVec2(simplex.c.p, simplex.b.p);

      Vec2 rab_dir = PerpVec2TowardsDir(vab, NegVec2(vac));
      Vec2 rac_dir = PerpVec2TowardsDir(vac, NegVec2(vab));
      Vec2 rbc_dir = PerpVec2TowardsDir(vbc, vab);

      f32 rab_dot = DotVec2(rab_dir, NegVec2(simplex.a.p));
      f32 rac_dot = DotVec2(rac_dir, NegVec2(simplex.a.p));
      f32 rbc_dot = DotVec2(rbc_dir, NegVec2(simplex.b.p));

      f32 vab_dot = DotVec2(vab, NegVec2(simplex.a.p)) / Vec2LenSq(vab);
      f32 vac_dot = DotVec2(vac, NegVec2(simplex.a.p)) / Vec2LenSq(vac);
      f32 vbc_dot = DotVec2(vbc, NegVec2(simplex.b.p)) / Vec2LenSq(vbc);

      if (rab_dot >= 0 && vab_dot >= 0 && vab_dot <= 1)
      {
        // Region ab, remove c
        num_removed = 1;
        removed_a = simplex.c.p;
        simplex.count = 2;
        dir = rab_dir;  // Next third point is in direction of region ab
      }
      else if (rac_dot >= 0 && vac_dot >= 0 && vac_dot <= 1)
      {
        // Region ac, remove b
        num_removed = 1;
        removed_a = simplex.b.p;
        simplex.count = 2;
        simplex.b = simplex.c;
        dir = rac_dir;  // Next third point is in direction of region ac
      }
      else if (rbc_dot >= 0 && vbc_dot >= 0 && vbc_dot <= 1)
      {
        // Region bc, remove a
        num_removed = 1;
        removed_a = simplex.a.p;
        simplex.count = 2;
        simplex.a = simplex.b;
        simplex.b = simplex.c;
        dir = rbc_dir;  // Next third point is in direction of region bc
      }
      else if (vab_dot <= 0 && vac_dot <= 0)
      {
        // Region a, remove bc
        num_removed = 2;
        removed_a = simplex.b.p;
        removed_b = simplex.c.p;
        simplex.count = 1;
      }
      else if (vab_dot >= 1 && vbc_dot <= 0)
      {
        // Region b, remove ac
        num_removed = 2;
        removed_a = simplex.a.p;
        removed_b = simplex.c.p;
        simplex.count = 1;
        simplex.a = simplex.b;
      }
      else if (vac_dot >= 1 && vbc_dot >= 1)
      {
        // Region c, remove ab
        num_removed = 2;
        removed_a = simplex.a.p;
        removed_b = simplex.b.p;
        simplex.count = 1;
        simplex.a = simplex.c;
      }
      else
      {
        // No region, must be in simplex
        is_overlapping = 1;
        break;
      }
    }
    if (!is_overlapping)
    {
      non_overlapping_dir = dir;
    }
  }

  //////////////////////////////
  //- EPA

  // Find dir from origin to closest edge
  Vec2 normal = Zi;
  P_MenkowskiSimplex closest_feature = Zi;
  {
    P_MenkowskiPoint *proto = 0;
    if (is_overlapping)
    {
      u32 proto_count = 0;
      proto = ArenaNext(scratch.arena, P_MenkowskiPoint);
      {
        Assert(simplex.count == 3);
        P_MenkowskiPoint *tmp = PushStructsNoZero(scratch.arena, P_MenkowskiPoint, 3);
        tmp[0] = simplex.a;
        tmp[1] = simplex.b;
        tmp[2] = simplex.c;
        proto_count = 3;
      }

      i32 winding = WindingFromVec2(SubVec2(simplex.c.p, simplex.a.p), SubVec2(simplex.b.p, simplex.a.p));

      u32 epa_iterations = 0;
      for (;;)
      {
        ++epa_iterations;

        // FIXME: Winding order of ps & pe index
        P_MenkowskiPoint closest_a = Zi;
        P_MenkowskiPoint closest_b = Zi;
        u32 closest_b_index = 0;
        {
          // Find edge segment on prototype closest to the origin
          f32 closest_len_sq = Inf;
          for (u32 i = 0; i < proto_count; ++i)
          {
            u32 a_index = i;
            u32 b_index = (i < proto_count - 1) ? (i + 1) : 0;
            P_MenkowskiPoint a = proto[a_index];
            P_MenkowskiPoint b = proto[b_index];
            Vec2 vab = SubVec2(b.p, a.p);
            Vec2 vao = NegVec2(a.p);
            f32 proj_ratio = ClampF32(DotVec2(vao, vab) / Vec2LenSq(vab), 0, 1);
            Vec2 proj = AddVec2(a.p, MulVec2(vab, proj_ratio));
            f32 proj_len_sq = Vec2LenSq(proj);
            if (proj_len_sq < closest_len_sq - min_unique_pt_dist_sq)
            {
              closest_a = a;
              closest_b = b;
              closest_b_index = b_index;
              closest_len_sq = proj_len_sq;
            }
          }
        }
        Vec2 vab = SubVec2(closest_b.p, closest_a.p);

        // Find new point in dir
        Vec2 dir = MulVec2(PerpVec2(vab), winding);
        P_MenkowskiPoint m = P_MenkowskiPointFromShapes(shape0, shape1, dir);

        // Check validity of new point
        {
          b32 valid = 1;
          {
            // NOTE: Changing this value affects how stable normals are for rounded colliders
            //const f32 validity_epsilon = min_unique_pt_dist_sq;  // Arbitrary
            //const f32 validity_epsilon = 0.00000000001f;  // Arbitrary
            const f32 validity_epsilon = min_unique_pt_dist_sq;  // Arbitrary

            Vec2 vam = SubVec2(m.p, closest_a.p);
            Vec2 vbm = SubVec2(closest_b.p, closest_a.p);

            f32 dot = DotVec2(vab, vam) / Vec2LenSq(vab);

            if (dot >= -validity_epsilon && dot <= 1 - validity_epsilon && (WedgeVec2(vab, vam) * -winding) >= -validity_epsilon)
            {
              // New point is not between edge
              valid = 0;
            }
            else if (Vec2LenSq(vam) < min_unique_pt_dist_sq || Vec2LenSq(vbm) < min_unique_pt_dist_sq)
            {
              // New point is too close to existing
              valid = 0;
            }
          }

          if (!valid || epa_iterations >= max_iterations)
          {
            normal = NormVec2(dir);
            closest_feature.a = closest_a;
            closest_feature.b = closest_b;
            closest_feature.count = 2;
            break;
          }
        }

        // Expand prototype
        PushStructNoZero(scratch.arena, P_MenkowskiPoint);
        ++proto_count;

        // Shift points in prototype to make room
        for (u32 i = proto_count - 1; i > closest_b_index; --i)
        {
          u32 shift_from = (i > 0) ? i - 1 : proto_count - 1;
          u32 shift_to = i;
          proto[shift_to] = proto[shift_from];
        }

        // Insert new point into prototype
        proto[closest_b_index] = m;
      }

      // Debug draw
      // {
      //   P_DebugDrawPoint(simplex.a.p, VEC4(1, 0, 0, 0.5));
      //   P_DebugDrawPoint(simplex.b.p, VEC4(0, 1, 0, 0.5));
      //   P_DebugDrawPoint(simplex.c.p, VEC4(0, 0, 1, 0.5));
      //   P_DebugDrawLine(simplex.a.p, simplex.b.p, Color_Yellow);
      //   P_DebugDrawLine(simplex.b.p, simplex.c.p, Color_Yellow);
      //   P_DebugDrawLine(simplex.c.p, simplex.a.p, Color_Yellow);
      //   if (proto_count > 0)
      //   {
      //     for (i64 i = 0; i < proto_count; ++i)
      //     {
      //       i64 p1_idx = i + 1;
      //       if (p1_idx == proto_count)
      //       {
      //         p1_idx = 0;
      //       }
      //       Vec2 p0 = proto[i].p;
      //       Vec2 p1 = proto[p1_idx].p;
      //       P_DebugDrawLine(p0, p1, VEC4(0, 1, 0, 0.5));
      //     }
      //   }
      // }
    }
    else
    {
      normal = NormVec2(non_overlapping_dir);
      closest_feature.count = simplex.count;
      closest_feature.a = simplex.a;
      closest_feature.b = simplex.b;
    }
  }

  //////////////////////////////
  //- Determine collision

  b32 is_colliding = 0;
  {
    if (is_overlapping)
    {
      is_colliding = 1;
    }
    else
    {
      // Shapes not overlapping, determine if distance between shapes within tolerance
      if (closest_feature.count == 1)
      {
        Vec2 p = NegVec2(closest_feature.a.p);
        if (Vec2LenSq(p) <= (tolerance * tolerance))
        {
          is_colliding = 1;
        }
      }
      else
      {
        // Project origin to determine if distance is within tolerance.
        Assert(closest_feature.count == 2);
        Vec2 vab = SubVec2(closest_feature.b.p, closest_feature.a.p);
        Vec2 vao = NegVec2(closest_feature.a.p);
        f32 ratio = ClampF32(DotVec2(vab, vao) / DotVec2(vab, vab), 0, 1);
        Vec2 p = AddVec2(closest_feature.a.p, MulVec2(vab, ratio));
        if (Vec2LenSq(p) <= (tolerance * tolerance))
        {
          is_colliding = 1;
        }
      }
    }
  }

  //////////////////////////////
  //- Compute collision points

  // Clip to determine final points
  i32 collision_points_count = 0;
  P_CollisionPoint collision_points[2] = Zi;
  if (is_colliding)
  {
    // Max vertices must be < 16 to fit in 4 bit ids
    StaticAssert(countof(shape0.points) <= 16);
    {
      b32 collapse0 = 0;
      b32 collapse1 = 0;

      P_SupportPoint a0 = closest_feature.a.s0;
      P_SupportPoint a1 = closest_feature.a.s1;
      P_SupportPoint b0 = closest_feature.b.s0;
      P_SupportPoint b1 = closest_feature.b.s1;
      // FIXME: Manually account for shapes w/ 1 & 2 points
      if (closest_feature.count == 2)
      {
        if (a0.id == b0.id)
        {
          if (shape0.points_count > 1)
          {
            b0 = P_SupportPointFromShapeEx(shape0, normal, b0.id);
          }
          else
          {
            collapse0 = 1;
            b0 = a0;
          }
        }
        if (a1.id == b1.id)
        {
          if (shape1.points_count > 1)
          {
            b1 = P_SupportPointFromShapeEx(shape1, NegVec2(normal), b1.id);
          }
          else
          {
            collapse1 = 1;
            b1 = a1;
          }
        }
      }
      else
      {
        collapse0 = 1;
        collapse1 = 1;
        b0 = a0;
        b1 = a1;
      }

      Vec2 vab0 = SubVec2(b0.p, a0.p);
      Vec2 vab1 = SubVec2(b1.p, a1.p);
      Vec2 vab0_norm = NormVec2(vab0);
      Vec2 vab1_norm = NormVec2(vab1);

      // Swap points based on normal direction for consistent clipping
      if (WedgeVec2(normal, vab0) < 0)
      {
        P_SupportPoint tmp = a0;
        a0 = b0;
        b0 = tmp;
        vab0 = NegVec2(vab0);
      }
      if (WedgeVec2(normal, vab1) < 0)
      {
        P_SupportPoint tmp = a1;
        a1 = b1;
        b1 = tmp;
        vab1 = NegVec2(vab1);
      }

      // Collapse lines that are too far in the direction of the normal to be accurately clipped
      f32 collapse_epsilon = 0.05f;
      collapse0 = collapse0 || AbsF32(WedgeVec2(normal, vab0_norm)) < collapse_epsilon;
      collapse1 = collapse1 || AbsF32(WedgeVec2(normal, vab1_norm)) < collapse_epsilon;

      // Collapse lines into deepest point
      if (collapse0)
      {
        if (DotVec2(normal, vab0) > 0)
        {
          a0 = b0;
        }
        else
        {
          // TODO: Remove this (debugging)
          b0 = a0;
        }
      }
      if (collapse1)
      {
        if (DotVec2(normal, vab1) < 0)
        {
          a1 = b1;
        }
        else
        {
          // TODO: Remove this (debugging)
          b1 = a1;
        }
      }

      f32 a_sep = Inf;
      f32 b_sep = Inf;
      Vec2 a_midpoint = Zi;
      Vec2 b_midpoint = Zi;
      b32 ignore_a = 1;
      b32 ignore_b = 1;
      if (!collapse0 && !collapse1)
      {
        // Clip line to line
        P_ClippedLine clip_result = P_ClipLineToLine(a0.p, b0.p, a1.p, b1.p, normal);
        Vec2 a0_clipped = clip_result.a0_clipped;
        Vec2 a1_clipped = clip_result.a1_clipped;
        Vec2 b0_clipped = clip_result.b0_clipped;
        Vec2 b1_clipped = clip_result.b1_clipped;
        // Calc midpoint between clipped a & b
        Vec2 va0a1_clipped = SubVec2(a1_clipped, a0_clipped);
        Vec2 vb0b1_clipped = SubVec2(b1_clipped, b0_clipped);
        a_sep = DotVec2(va0a1_clipped, normal);
        b_sep = DotVec2(vb0b1_clipped, normal);
        a_midpoint = AddVec2(a0_clipped, MulVec2(va0a1_clipped, 0.5f));
        b_midpoint = AddVec2(b0_clipped, MulVec2(vb0b1_clipped, 0.5f));
        ignore_a = 0;
        ignore_b = 0;
        Vec2 vfin = SubVec2(b_midpoint, a_midpoint);
        if (Vec2LenSq(vfin) < (0.005 * 0.005))
        {
          if (a_sep > b_sep)
          {
            ignore_a = 1;
          }
          else
          {
            ignore_b = 1;
          }
        }
      }
      else
      {
        Vec2 p0 = a0.p;
        Vec2 p1 = a1.p;
        // TODO: Choose ID based on closest clipped point
        if (collapse1 && !collapse0)
        {
          // Project a1 onto vab0
          p0 = P_ClipPointToLine(a0.p, b0.p, a1.p, normal);
        }
        if (collapse0 && !collapse1)
        {
          // Project a0 onto vab1
          p1 = P_ClipPointToLine(a1.p, b1.p, a0.p, normal);
        }
        // Calc midpoint
        Vec2 vsep = SubVec2(p1, p0);
        a_midpoint = AddVec2(p0, MulVec2(vsep, 0.5f));
        a_sep = DotVec2(normal, p1) - DotVec2(normal, p0);
        ignore_a = 0;
      }

      // Insert points
      if (!ignore_a && a_sep < tolerance)
      {
        P_CollisionPoint *point = &collision_points[collision_points_count++];
        point->id = a0.id | (a1.id << 4);
        point->separation = a_sep;
        point->p = a_midpoint;
      }
      if (!ignore_b && b_sep < tolerance)
      {
        P_CollisionPoint *point = &collision_points[collision_points_count++];
        point->id = b0.id | (b1.id << 4);
        point->separation = b_sep;
        point->p = b_midpoint;
      }
    }
  }

  //////////////////////////////
  //- Compute closest points

  Vec2 closest_p0 = Zi;
  Vec2 closest_p1 = Zi;
  if (closest_feature.count == 1)
  {
    closest_p0 = closest_feature.a.s0.p;
    closest_p1 = closest_feature.a.s1.p;
  }
  else
  {
    Assert(closest_feature.count == 2);
    // FIXME: Winding order dependent?
    f32 ratio = 0;
    {
      // Determine ratio between edge a & b that projected origin lies
      Vec2 vab = SubVec2(closest_feature.b.p, closest_feature.a.p);
      Vec2 vao = NegVec2(closest_feature.a.p);
      ratio = ClampF32(DotVec2(vab, vao) / DotVec2(vab, vab), 0, 1);
    }
    // Shape 0
    closest_p0 = SubVec2(closest_feature.b.s0.p, closest_feature.a.s0.p);
    closest_p0 = MulVec2(closest_p0, ratio);
    closest_p0 = AddVec2(closest_p0, closest_feature.a.s0.p);
    // Shape 1
    closest_p1 = SubVec2(closest_feature.b.s1.p, closest_feature.a.s1.p);
    closest_p1 = MulVec2(closest_p1, ratio);
    closest_p1 = AddVec2(closest_p1, closest_feature.a.s1.p);
  }

  CopyStructs(result.collision_points, collision_points, countof(collision_points));
  result.collision_points_count = collision_points_count;
  result.collision_normal = normal;
  result.closest_p0 = closest_p0;
  result.closest_p1 = closest_p1;

  EndScratch(scratch);
  return result;
}

P_RaycastResult P_RaycastShape(P_Shape shape, Vec2 ray_start, Vec2 ray_dir)
{
  f32 radius = shape.radius;

  f32 ray_len = Vec2Len(ray_dir);
  Vec2 ray_dir_norm = DivVec2(ray_dir, ray_len);

  Vec2 s = ray_start;
  Vec2 e = AddVec2(ray_start, ray_dir_norm);
  Vec2 vse = ray_dir_norm;

  Vec2 isect = Zi;
  Vec2 isect_normal = Zi;
  b32 isect_is_round = 0;
  b32 isect_found = 0;
  if (shape.points_count > 1)
  {
    // Find expanded line intersections with ray
    for (i32 p_idx = 0; p_idx < shape.points_count && !isect_found; ++p_idx)
    {
      Vec2 a = Zi;
      Vec2 b = Zi;
      Vec2 vab = Zi;
      Vec2 normal = Zi;
      {
        i32 a_idx = p_idx;
        i32 b_idx = a_idx + 1;
        if (b_idx >= shape.points_count)
        {
          b_idx = 0;
        }

        Vec2 a_orig = shape.points[a_idx];
        Vec2 b_orig = shape.points[b_idx];
        vab = SubVec2(b_orig, a_orig);
        normal = NegVec2(PerpVec2(NormVec2(vab)));
        Vec2 radius_add = MulVec2(normal, radius);

        a = AddVec2(a_orig, radius_add);
        b = AddVec2(b_orig, radius_add);
      }
      Vec2 vsa = SubVec2(a, s);
      Vec2 vsb = SubVec2(b, s);
      f32 wa = WedgeVec2(vse, vsa);
      f32 wb = WedgeVec2(vse, vsb);
      if (wa > 0 && wb < 0)
      {
        f32 t = -wa / (wb - wa);
        isect = AddVec2(a, MulVec2(vab, t));
        isect_normal = normal;
        isect_found = 1;
      }
    }

    // Find closest rounded corner
    if (!isect_found && radius != 0)
    {
      isect_is_round = 1;
      for (i32 f_idx = 0; f_idx < shape.points_count && !isect_found; ++f_idx)
      {
        Vec2 f_orig = shape.points[f_idx];
        Vec2 a = Zi;
        Vec2 b = Zi;
        Vec2 vab = Zi;
        {
          i32 prev_idx = f_idx - 1;
          i32 next_idx = f_idx + 1;
          if (prev_idx < 0)
          {
            prev_idx = shape.points_count - 1;
          }
          if (next_idx >= shape.points_count)
          {
            next_idx = 0;
          }
          Vec2 prev_orig = shape.points[prev_idx];
          Vec2 next_orig = shape.points[next_idx];
          Vec2 vpf = SubVec2(f_orig, prev_orig);
          Vec2 vfn = SubVec2(next_orig, f_orig);
          Vec2 vpf_norm = NormVec2(vpf);
          Vec2 vfn_norm = NormVec2(vfn);
          Vec2 radius_add_a = MulVec2(PerpVec2(vpf_norm), -radius);
          Vec2 radius_add_b = MulVec2(PerpVec2(vfn_norm), -radius);
          a = AddVec2(f_orig, radius_add_a);
          b = AddVec2(f_orig, radius_add_b);
        }
        Vec2 vsa = SubVec2(a, s);
        Vec2 vsb = SubVec2(b, s);
        f32 wa = WedgeVec2(vse, vsa);
        f32 wb = WedgeVec2(vse, vsb);
        if (wa > 0 && wb < 0)
        {
          isect = f_orig;
          isect_found = 1;
        }
      }
    }

    // Find closest corner
    if (!isect_found)
    {
      f32 min_dist = Inf;
      for (i32 p_idx = 0; p_idx < shape.points_count && !isect_found; ++p_idx)
      {
        Vec2 p = shape.points[p_idx];
        f32 dist = AbsF32(WedgeVec2(vse, SubVec2(p, s)));
        if (dist < min_dist)
        {
          isect = p;
          min_dist = dist;
        }
      }
    }
  }
  else if (shape.points_count == 1 && radius != 0)
  {
    isect = shape.points[0];
    isect_is_round = 1;
  }


  // Find round intersection
  b32 is_intersecting = 0;
  if (isect_is_round || !isect_found)
  {
    Vec2 vsi = SubVec2(isect, s);

    f32 dot = DotVec2(vse, vsi);
    f32 wedge = WedgeVec2(vse, vsi);

    is_intersecting = AbsF32(wedge) < radius;
    if (is_intersecting)
    {
      f32 diff = SqrtF32(radius * radius - wedge * wedge);
      f32 entrance_t = dot - diff;
      f32 exit_t = dot + diff;
      {
        Vec2 old_isect = isect;
        isect = AddVec2(s, MulVec2(vse, entrance_t));
        isect_normal = NormVec2(SubVec2(isect, old_isect));
      }
    }
  }
  else
  {
    is_intersecting = isect_found;
  }

  P_RaycastResult result = Zi;
  result.is_intersecting = is_intersecting;
  if (is_intersecting)
  {
    result.p = isect;
    result.normal = isect_normal;
  }

  return result;
}

Vec2 P_EdgePointFromShape(P_Shape shape, Vec2 dir)
{
  Vec2 result = shape.centroid;
  P_RaycastResult raycast = P_RaycastShape(shape, shape.centroid, NegVec2(dir));
  if (raycast.is_intersecting)
  {
    result = raycast.p;
  }
  return result;
}

////////////////////////////////////////////////////////////
//~ Lookup helpers

P_Ent *P_EntFromKey(P_Frame *frame, P_Key key)
{
  P_Ent *result = &P_NilEnt;
  P_World *world = frame->world;
  if (!P_IsKeyNil(key) && frame->tick > 0 && frame->ents_count > 0 && frame->ent_bins_count > 0)
  {
    i64 tick = frame->tick;
    P_EntBin *bin = &frame->ent_bins[key.v % frame->ent_bins_count];
    for (P_Ent *e = bin->first; e; e = e->next_in_bin)
    {
      if (e->key.v == key.v)
      {
        result = e;
        break;
      }
    }
  }
  return result;
}

////////////////////////////////////////////////////////////
//~ Iteration helpers

P_Ent *P_FirstEnt(P_Frame *frame)
{
  P_Ent *result = &P_NilEnt;
  if (!P_IsEntNil(frame->first_ent))
  {
    result = frame->first_ent;
  }
  return result;
}

P_Ent *P_NextEnt(P_Ent *e)
{
  P_Ent *result = &P_NilEnt;
  if (!P_IsEntNil(e) && !P_IsEntNil(e->next))
  {
    result = e->next;
  }
  return result;
}

P_Ent *P_PushTempEnt(Arena *arena, P_EntList *list)
{
  P_EntListNode *n = PushStruct(arena, P_EntListNode);
  SllQueuePush(list->first, list->last, n);
  ++list->count;
  P_Ent *ent = &n->ent;
  *ent = P_NilEnt;
  ent->exists = 1;
  return ent;
}

////////////////////////////////////////////////////////////
//~ Debug draw

void P_DebugDrawPoint(Vec2 p, Vec4 srgb)
{
  if (P_tl.debug_draw_enabled)
  {
    P_DebugDrawNode *n = PushStruct(P_tl.debug_arena, P_DebugDrawNode);
    {
      n->kind = P_DebugDrawKind_Point;
      n->srgb32 = U32FromVec4(srgb);
      n->point.p = p;
    }
    SllQueuePush(P_tl.first_debug_draw_node, P_tl.last_debug_draw_node, n);
    P_tl.debug_draw_nodes_count += 1;
  }
}

void P_DebugDrawLine(Vec2 p0, Vec2 p1, Vec4 srgb)
{
  if (P_tl.debug_draw_enabled)
  {
    P_DebugDrawNode *n = PushStruct(P_tl.debug_arena, P_DebugDrawNode);
    {
      n->kind = P_DebugDrawKind_Line;
      n->srgb32 = U32FromVec4(srgb);
      n->line.p0 = p0;
      n->line.p1 = p1;
    }
    SllQueuePush(P_tl.first_debug_draw_node, P_tl.last_debug_draw_node, n);
    P_tl.debug_draw_nodes_count += 1;
  }
}

void P_DebugDrawRect(Rng2 rect, Vec4 srgb)
{
  if (P_tl.debug_draw_enabled)
  {
    P_DebugDrawNode *n = PushStruct(P_tl.debug_arena, P_DebugDrawNode);
    {
      n->kind = P_DebugDrawKind_Rect;
      n->srgb32 = U32FromVec4(srgb);
      n->rect = rect;
    }
    SllQueuePush(P_tl.first_debug_draw_node, P_tl.last_debug_draw_node, n);
    P_tl.debug_draw_nodes_count += 1;
  }
}

void P_DebugDrawShape(P_Shape shape, Vec4 srgb)
{
  if (P_tl.debug_draw_enabled)
  {
    P_DebugDrawNode *n = PushStruct(P_tl.debug_arena, P_DebugDrawNode);
    {
      n->kind = P_DebugDrawKind_Shape;
      n->srgb32 = U32FromVec4(srgb);
      n->shape = shape;
    }
    SllQueuePush(P_tl.first_debug_draw_node, P_tl.last_debug_draw_node, n);
    P_tl.debug_draw_nodes_count += 1;
  }
}

////////////////////////////////////////////////////////////
//~ World

P_World *P_AcquireWorld(void)
{
  P_World *world = 0;
  {
    Arena *arena = AcquireArena(Gibi(64));
    world = PushStruct(arena, P_World);
    world->arena = arena;
  }
  world->frames_arena = AcquireArena(Gibi(64));
  world->statics_arena = AcquireArena(Gibi(64));

  world->first_frame = &P_NilFrame;
  world->last_frame = &P_NilFrame;
  world->frame_bins_count = Kibi(16);
  world->frame_bins = PushStructs(world->arena, P_FrameBin, world->frame_bins_count);

  // TODO

  world->tiles = PushStructs(world->arena, u8, P_TilesCount);

  TrueRand(StringFromStruct(&world->seed));
  return world;
}

void P_SpawnEntsFromList(P_Frame *frame, P_EntList ents)
{
  P_World *world = frame->world;
  for (P_EntListNode *n = ents.first; n; n = n->next)
  {
    P_Ent *src = &n->ent;
    P_Key key = src->key;
    if (!P_IsKeyNil(src->key))
    {
      P_EntBin *bin = &frame->ent_bins[key.v % frame->ent_bins_count];
      P_Ent *dst = bin->first;
      for (; dst; dst = dst->next_in_bin)
      {
        if (dst->key.v == key.v)
        {
          break;
        }
      }
      if (!dst)
      {
        dst = world->first_free_ent;
        if (dst)
        {
          SllStackPop(world->first_free_ent);
        }
        else
        {
          dst = PushStructNoZero(world->frames_arena, P_Ent);
        }
        DllQueuePushNPZ(&P_NilEnt, frame->first_ent, frame->last_ent, dst, next, prev);
        DllQueuePushNP(bin->first, bin->last, dst, next_in_bin, prev_in_bin);
      }
      P_Ent *old_next = dst->next;
      P_Ent *old_prev = dst->prev;
      P_Ent *old_next_in_bin = dst->next_in_bin;
      P_Ent *old_prev_in_bin = dst->prev_in_bin;
      {
        *dst = *src;
      }
      dst->next = old_next;
      dst->prev = old_prev;
      dst->next_in_bin = old_next_in_bin;
      dst->prev_in_bin = old_prev_in_bin;
      ++frame->ents_count;
    }
  }
}

P_Frame *P_FrameFromTick(P_World *world, i64 tick)
{
  P_Frame *result = &P_NilFrame;
  if (world->frame_bins_count > 0)
  {
    u64 hash = MixU64(tick);
    P_FrameBin *bin = &world->frame_bins[hash % world->frame_bins_count];
    for (P_Frame *frame = bin->first; frame; frame = frame->next_in_bin)
    {
      if (frame->tick == tick)
      {
        result = frame;
        break;
      }
    }
  }
  return result;
}

void P_ClearFrames(P_World *world, i64 tick_min, i64 tick_max)
{
  // TODO: Fast path for when range encompasses all frames in the world
  // TODO: Don't need linear search
  P_Frame *frame = world->first_frame;
  while (!P_IsFrameNil(frame))
  {
    P_Frame *next_frame = frame->next;
    if (frame->tick >= tick_min && frame->tick <= tick_max)
    {
      if (!P_IsEntNil(frame->first_ent))
      {
        frame->last_ent->next = world->first_free_ent;
        world->first_free_ent = frame->first_ent;
      }

      u64 hash = MixU64(frame->tick);
      P_FrameBin *bin = &world->frame_bins[hash % world->frame_bins_count];
      DllQueueRemoveNPZ(&P_NilFrame, world->first_frame, world->last_frame, frame, next, prev);
      DllQueueRemoveNPZ(0, bin->first, bin->last, frame, next_in_bin, prev_in_bin);

      SllStackPush(world->first_free_frame, frame);
    }
    else
    {
      break;
    }
    frame = next_frame;
  }
}

P_Frame *P_PushFrame(P_World *world, P_Frame *src_frame, i64 tick)
{
  Assert(!(src_frame->world == world && tick <= src_frame->tick));  // Can't read from tick that is being overwritten by new tick
  P_ClearFrames(world, tick, I64Max);

  P_Frame *frame = world->first_free_frame;
  if (frame)
  {
    SllStackPop(world->first_free_frame);
    i64 old_ent_bins_count = frame->ent_bins_count;
    P_EntBin *old_ent_bins = frame->ent_bins;
    ZeroStruct(frame);
    frame->ent_bins_count = old_ent_bins_count;
    frame->ent_bins = old_ent_bins;
    ZeroStructs(frame->ent_bins, frame->ent_bins_count);
  }
  else
  {
    frame = PushStruct(world->frames_arena, P_Frame);
  }

  {
    frame->world = world;
    frame->tick = tick;
    frame->time_ns = src_frame->time_ns;

    frame->first_ent = &P_NilEnt;
    frame->last_ent = &P_NilEnt;
    if (frame->ent_bins_count == 0)
    {
      frame->ent_bins_count = Kibi(16);
      frame->ent_bins = PushStructs(world->frames_arena, P_EntBin, frame->ent_bins_count);
    }

    u64 hash = MixU64(tick);
    P_FrameBin *bin = &world->frame_bins[hash % world->frame_bins_count];

    DllQueuePushNPZ(&P_NilFrame, world->first_frame, world->last_frame, frame, next, prev);
    DllQueuePushNPZ(0, bin->first, bin->last, frame, next_in_bin, prev_in_bin);
  }

  for (P_Ent *src = P_FirstEnt(src_frame); !P_IsEntNil(src); src = P_NextEnt(src))
  {
    // FIXME: Pull from freelist
    P_Ent *dst = world->first_free_ent;
    if (dst)
    {
      SllStackPop(world->first_free_ent);
    }
    else
    {
      dst = PushStructNoZero(world->frames_arena, P_Ent);
    }
    *dst = *src;
    P_EntBin *bin = &frame->ent_bins[src->key.v % frame->ent_bins_count];
    DllQueuePushNPZ(&P_NilEnt, frame->first_ent, frame->last_ent, dst, next, prev);
    DllQueuePushNP(bin->first, bin->last, dst, next_in_bin, prev_in_bin);
    ++frame->ents_count;
  }

  return frame;
}

////////////////////////////////////////////////////////////
//~ Step

P_Frame *P_StepWorld(P_World *world, P_Frame *prev_frame, P_CmdList cmds)
{
  P_Frame *result = &P_NilFrame;
  return result;
}